🧠 Modèle prédictif
P(open|user,t) = Σᵢ wᵢ × fᵢ(user, t)
Features:
f₁ = timezone_match(user.tz, send_time) w=0.25
f₂ = historical_open_hour(user, hour) w=0.30
f₃ = day_of_week_preference(user, dow) w=0.15
f₄ = device_activity_pattern(user, t) w=0.15
f₅ = isp_congestion(isp, t) w=0.10
f₆ = recency_decay(last_open, t) w=0.05
Training: 847K profils × 14 jours = 11.8M datapoints